Online monitoring system for auxiliary transformer of heavy haul railway locomotive

Abstract

The invention discloses a heavy haul railway locomotive auxiliary transformer on-line monitoring system which comprises a sensing system, a control system and a monitoring center, the sensing system is in communication connection with the control system, and the control system is in communication connection with the monitoring center through a wireless communication network. According to the on-line monitoring system for the auxiliary transformer of the heavy haul railway locomotive, through cooperative work of the sensing system, the control system and the monitoring center, real-time acquisition, transmission, analysis and remote monitoring of core operation parameters of the auxiliary transformer are realized, and the problem that the equipment state cannot be mastered in real time through traditional regular maintenance and off-line detection is solved; the 5G wireless transmission and encryption technology guarantees the stability and safety of data transmission, the multi-channel high-speed acquisition module guarantees the data accuracy, the data sharing and report function of the monitoring center provides basic decision support for operation and maintenance, the safety and reliability of operation of the heavy haul railway locomotive auxiliary transformer are remarkably improved, and the safety and reliability of the heavy haul railway locomotive auxiliary transformer are improved. And the sudden fault risk is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the online monitoring system technical field, specifically to a kind of heavy haul railway locomotive auxiliary transformer online monitoring system. BACKGROUND

[0002] Heavy haul railway locomotive auxiliary transformer assumes important task in the process of running, such as traction fan, cooling tower fan and auxiliary transformer cabinet fan, its working condition directly influences the normal operation and safety performance of locomotive, and the traditional transformer monitoring mode is mostly periodic maintenance and offline detection, cannot master the real-time operation state of transformer, and it is difficult to find potential fault in time. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a kind of heavy haul railway locomotive auxiliary transformer online monitoring system, solve the problems raised in the above background art.

[0004] To achieve the above object, the present application is realized by the following technical scheme: a kind of heavy haul railway locomotive auxiliary transformer online monitoring system, including sensing system, control system and monitoring center, the sensing system is connected with control system communication, the control system is connected with monitoring center by wireless communication network communication; The sensing system includes temperature sensor, thermal imaging sensor, voltage sensor, current sensor, power quality monitoring sensor, resistance measuring sensor and insulation resistance sensor, and sensing system integrates data acquisition module and data transmission module, the data acquisition module has multi-channel synchronous acquisition, high-speed sampling and data cache function; The control system includes data analysis module and fault early warning module, the data analysis module can realize data preprocessing, threshold judgment, trend analysis, fault diagnosis and data mining function, the fault early warning module can send early warning signal by sound, light, SMS mode and record fault related information; The monitoring center includes remote monitoring module and data storage and management module, the remote monitoring module supports remote access, control and debugging, the data storage and management module uses MySQL or Oracle database, has data backup, recovery, statistics and report generation function.

[0005] Further, the wireless communication network uses 5G wireless transmission technology, and the data transmission process uses encryption technology processing.

[0006] Further, the sampling frequency of the data acquisition module is adapted to auxiliary transformer operation parameter monitoring demand.

[0007] Furthermore, the data analysis module integrates an adaptive fusion learning unit based on Attention-LSTM+XGBoost.

[0008] Furthermore, the adaptive fusion learning unit extracts key parameter features through an attention mechanism, constructs a multi-model collaborative diagnostic framework by combining it with the gradient boosting tree algorithm, and realizes quantitative assessment of the equipment's full life cycle based on the Remaining Life (RUL) prediction algorithm.

[0009] Furthermore, the training dataset of the adaptive fusion learning unit includes the factory rated parameters of transformers of the same model, individual historical operation-fault data, and real-time operation parameters. The remaining life prediction results are output in a three-dimensional mode, and the health level corresponds to a clear parameter threshold range. The life expectancy prediction results are output in a three-dimensional format as follows: quantified value + health level + decline rate. The quantified value is in monthly units, the health level includes excellent, good, moderate, poor, and critical, and the decline rate is in % / month.

[0010] Furthermore, the adaptive fusion learning unit adopts a hierarchical iterative optimization mechanism associated with health levels. Different health levels correspond to different iteration triggering conditions, and the iteration process incorporates manual verification data feedback calibration, with calibration accounting for 30% of the weight.

[0011] Furthermore, the data analysis module also integrates an intelligent verification unit for abnormal data, which filters invalid data through a triple mechanism of parameter correlation verification, sensor self-diagnosis, and historical data comparison.

[0012] Furthermore, the specific implementation process of the adaptive fusion learning unit is as follows: Based on LSTM neural networks, an attention mechanism is introduced to automatically assign parameter weights, focusing on parameters that play a key role in fault diagnosis and life prediction, including winding temperature and insulation resistance. The specific feature extraction formula is as follows: Hidden layer state update:

[0013] Attention weight calculation:

[0014] Weighted feature output:

[0015] in, for Input parameter vector at any time, This represents the hidden state of the LSTM layer. For attention weight vectors, The final extracted key feature vector, , , , Parameters for model training; By integrating the XGBoost gradient boosting tree algorithm, a multi-model collaborative framework is constructed, and diagnostic results are output through a weighted voting mechanism, as shown in the following formula: Model output fusion:

[0016] in, For the final fault diagnosis results, , These are the diagnostic outputs for the LSTM and XGBoost models, respectively. These are weighting coefficients; A quantitative model is constructed based on parameter decay patterns and historical fault data, with the following formula: Remaining lifespan base value:

[0017] Corrected remaining lifetime:

[0018] in, For the parameter, allowable threshold for Measured values ​​of time parameters This represents the average decay rate over the past three months. For safety reasons, This is the load factor for the operating condition, with a value between 0 and 1. This is the environmental impact coefficient, with a value between 0 and 1. , These are the weighting coefficients; The remaining life expectancy prediction results are output in three dimensions using a format of quantitative values ​​+ health level + decline rate.

[0019] Furthermore, the model iteration strategy is dynamically adjusted based on the equipment health level: A health rating of Excellent or Good is achieved by accumulating 300 hours of valid data or triggering an iteration due to one instance of abnormal parameter fluctuation. Health level is medium or poor: 50 hours of valid data or two consecutive iterations triggered by unrelated anomalies; Health level is critical: Model parameters are updated iteratively in real time; During the iteration process, manually entered maintenance records and fault verification results are weighted at 30% for calibration, as shown in the following formula:

[0020] in, For the updated model parameters, For iterative calculation of parameters, The calibration parameters were manually verified. The intelligent verification and deployment optimization of abnormal data adopts a triple verification mechanism, as follows: Parameter correlation verification: Parameter consistency is judged based on the Pearson correlation coefficient, as shown in the following formula:

[0021] The association is deemed valid; Sensor self-diagnosis: Monitoring sensor signal stability and coefficient of variation The sensor was determined to be normal. Indicates the sample standard deviation. This represents the sample mean; Historical data comparison: Calculate the deviation rate between the current data and the historical data for the same period. If:

[0022] Then it is determined to be valid, where, This is the average value of the parameter over a historical period.

[0023] This invention provides an online monitoring system for auxiliary transformers of heavy-haul railway locomotives, which has the following advantages: 1. This online monitoring system for auxiliary transformers of heavy-haul railway locomotives, through the collaborative work of the sensing system, control system, and monitoring center, realizes the real-time acquisition, transmission, analysis, and remote monitoring of the core operating parameters of the auxiliary transformers. It solves the problem that traditional periodic maintenance and offline testing cannot grasp the equipment status in real time. 5G wireless transmission and encryption technology ensure the stability and security of data transmission, multi-channel high-speed acquisition modules ensure data accuracy, and the data sharing and reporting functions of the monitoring center provide basic decision support for operation and maintenance. It significantly improves the safety and reliability of the operation of auxiliary transformers of heavy-haul railway locomotives and reduces the risk of sudden failures.

[0024] 2. This online monitoring system for auxiliary transformers of heavy-haul railway locomotives accurately extracts key features by combining Attention-LSTM+XGBoost multi-model fusion with an attention mechanism. It also improves the accuracy of fault diagnosis and reduces the false alarm rate by using triple anomaly data verification. This solves the problems of poor adaptability and high misjudgment rate of single algorithms. Furthermore, the RUL algorithm integrates multi-dimensional influencing factors, and the three-dimensional output mode clearly presents the equipment degradation status, providing solid data support for predictive maintenance. Attached Figure Description

[0025] Fig. 1 This is a flowchart illustrating the steps of the sensing system, control system, and monitoring center of an online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to the present invention. Fig. 2This is a schematic diagram of the adaptive fusion learning unit steps of an online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to the present invention. Detailed Implementation

[0026] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0027] like Figs. 1-2 As shown, the present invention provides a technical solution: an online monitoring system for auxiliary transformers of heavy-haul railway locomotives, comprising a sensing system, a control system, and a monitoring center, wherein the sensing system is communicatively connected to the control system, and the control system is communicatively connected to the monitoring center through a wireless communication network; The sensing system includes a temperature sensor, a thermal imaging sensor, a voltage sensor, a current sensor, a power quality monitoring sensor, a resistance measurement sensor, and an insulation resistance sensor. The sensing system integrates a data acquisition module and a data transmission module. The data acquisition module has multi-channel synchronous acquisition, high-speed sampling, and data caching functions. The control system includes a data analysis module and a fault early warning module. The data analysis module can perform data preprocessing, threshold judgment, trend analysis, fault diagnosis and data mining. The fault early warning module can issue early warning signals and record fault-related information through sound, light and text messages. The monitoring center includes a remote monitoring module and a data storage and management module. The remote monitoring module supports remote access, control and debugging. The data storage and management module uses a MySQL or Oracle database and has data backup, recovery, statistics and report generation functions. The wireless communication network uses 5G wireless transmission technology, and the data transmission process is encrypted. The sampling frequency of the data acquisition module is adapted to the monitoring requirements of the auxiliary transformer's operating parameters; Among them, the temperature sensor is embedded in the key part of the winding, the thermal imaging sensor is installed on the top of the cabinet, the voltage and current sensors are connected in series or parallel to the power supply circuit, and the remaining sensors are installed at the insulation detection point and the winding resistance detection end respectively; the data acquisition module has a sampling frequency of 1kHz-10kHz and a buffer capacity of ≥16GB; the 5G transmission module uses encryption technology to process the transmitted data. The data analysis module preprocesses the collected data and performs basic fault diagnosis. The fault early warning module issues early warnings and records fault information according to a preset method. The data is synchronously stored in the local database and uploaded to the monitoring center. The remote monitoring module supports B / S architecture access. The data storage and management module stores data according to a daily automatic backup mechanism, and the backup is retained for 3 months. The report generation function supports the output of basic operation reports on a periodic basis. Example 1: Sensor System Deployment: PT100 temperature sensors are embedded in three key locations of the auxiliary transformer for heavy-haul railway locomotives: the middle of the high-voltage winding and the end of the low-voltage winding. A FLIR A315 thermal imaging sensor is installed on the top of the cabinet to monitor the surface temperature distribution in real time. A high-precision current transformer (rated current 5A, accuracy 0.2S class) is connected in series in the primary power supply circuit of the transformer, and a voltage transformer (turn ratio 10000 / 100V, accuracy 0.5 class) is connected in parallel on the secondary side. At the same time, a partial discharge sensor and a DC resistance tester are installed at the insulating oil sampling port and the winding lead terminals, respectively. The data acquisition module uses Advantech ADAM-6017, with a sampling frequency set to 5kHz, and a 256GB industrial-grade SSD is configured as a cache. The 5G transmission module uses Huawei ME909s-821, and data security is achieved through the AES-256 encryption protocol. Control system operation: When the freight train travels to the Northwest Gobi Desert section, the sensor detects that the temperature of the transformer phase winding rises sharply from 65°C to 85°C within 2 hours, exceeding the 75°C warning threshold. The data analysis module is immediately activated, performs moving average filtering on the temperature data, and constructs a temperature trend model by combining it with historical operating data. It determines that there is a risk of cooling system failure. The fault warning module flashes a red warning light on the on-board display and issues a voice alarm to the driver. At the same time, it records the fault time, type, and related data to the local SQLite database and uploads it to the monitoring center via the 5G network. Monitoring Center Management: At the electric locomotive dispatch center, maintenance engineers access the remote monitoring system with a B / S architecture via a web browser to view the real-time operating status of the train's auxiliary transformer. The data storage and management module automatically compresses and backs up the data of the previous 24 hours to the NAS storage array at 2:00 AM every day, retaining the data of the most recent 90 days. Based on maintenance needs, the system automatically generates operation analysis reports every Monday, including load rate, temperature curve, fault statistics, etc., for technicians to assess the health status of the equipment. Based on the above description, this invention achieves real-time acquisition, transmission, analysis, and remote monitoring of core operating parameters of auxiliary transformers through the collaborative work of the sensing system, control system, and monitoring center. It solves the problem that traditional periodic maintenance and offline testing cannot grasp the equipment status in real time. 5G wireless transmission and encryption technology ensure the stability and security of data transmission, multi-channel high-speed acquisition modules ensure data accuracy, and the data sharing and reporting functions of the monitoring center provide basic decision support for operation and maintenance. This significantly improves the safety and reliability of the operation of auxiliary transformers for heavy-haul railway locomotives and reduces the risk of sudden failures. The data analysis module integrates an adaptive fusion learning unit based on Attention-LSTM+XGBoost. The adaptive fusion learning unit extracts key parameter features through the attention mechanism and constructs a multi-model collaborative diagnostic framework by combining the gradient boosting tree algorithm. At the same time, it realizes the quantitative assessment of the equipment's full life cycle based on the Remaining Life (RUL) prediction algorithm. The training dataset of the adaptive fusion learning unit includes the factory rated parameters of transformers of the same model, individual historical operation-fault data, and real-time operation parameters. The remaining life prediction results are output in a three-dimensional mode, and the health level corresponds to a clear parameter threshold range. The life prediction results are output in a three-dimensional mode in the following format: quantified value + health level + decline rate. The quantified value is in monthly units, the health level includes excellent, good, moderate, poor, and critical, and the decline rate is in % / month. Quantitative values: in months, with an accuracy of ±0.5 months;

[0028] Health levels correspond to specific parameter thresholds, such as: Advantages: Insulation resistance ≥500MΩ, winding resistance attenuation ≤5%, temperature fluctuation ≤±3℃; Good: 300MΩ ≤ insulation resistance < 500MΩ, 5% < winding resistance attenuation ≤ 10%; Medium: 100MΩ ≤ insulation resistance < 300MΩ, 10% < winding resistance attenuation ≤ 15%; Poor insulation resistance: 50MΩ ≤ insulation resistance < 100MΩ, 15% < winding resistance attenuation ≤ 20%; Danger: Insulation resistance <50MΩ, winding resistance attenuation >20%; Decline rate: Unit "% / month", reflecting the trend of parameter decay speed; The adaptive fusion learning unit adopts a hierarchical iterative optimization mechanism based on health level correlation. Different health levels correspond to different iteration trigger conditions, and the iteration process incorporates manual verification data feedback calibration, with calibration accounting for 30% of the weight. The data analysis module also integrates an intelligent verification unit for abnormal data, which filters invalid data through a triple mechanism of parameter correlation verification, sensor self-diagnosis, and historical data comparison. The specific implementation process of the adaptive fusion learning unit is as follows: Based on LSTM neural networks, an attention mechanism is introduced to automatically assign parameter weights, focusing on parameters that play a key role in fault diagnosis and life prediction, including winding temperature and insulation resistance. The specific feature extraction formula is as follows: Hidden layer state update:

[0029] Attention weight calculation:

[0030] Weighted feature output:

[0031] in, for Input parameter vector at any time, This represents the hidden state of the LSTM layer. For attention weight vectors, The final extracted key feature vector, , , , Parameters for model training; By integrating the XGBoost gradient boosting tree algorithm, a multi-model collaborative framework is constructed, and diagnostic results are output through a weighted voting mechanism, as shown in the following formula: Model output fusion:

[0032] in, For the final fault diagnosis results, , These are the diagnostic outputs for the LSTM and XGBoost models, respectively. These are weighting coefficients; A quantitative model is constructed based on parameter decay patterns and historical fault data, with the following formula: Remaining lifespan base value:

[0033] Corrected remaining lifetime:

[0034] in, For the parameter, allowable threshold for Measured values ​​of time parameters This represents the average decay rate over the past three months. For safety reasons, This is the load factor for the operating condition, with a value between 0 and 1. This is the environmental impact coefficient, with a value between 0 and 1. , These are the weighting coefficients; The remaining life prediction results are output in three dimensions using a format of quantified numerical values ​​+ health level + degradation rate. The model iteration strategy is dynamically adjusted based on the equipment health level. A health rating of Excellent or Good is achieved by accumulating 300 hours of valid data or triggering an iteration due to one instance of abnormal parameter fluctuation. Health level is medium or poor: 50 hours of valid data or two consecutive iterations triggered by unrelated anomalies; Health level is critical: Model parameters are updated iteratively in real time; During the iteration process, manually entered maintenance records and fault verification results are weighted at 30% for calibration, as shown in the following formula:

[0035] in, For the updated model parameters, For iterative calculation of parameters, The calibration parameters were manually verified. The intelligent verification and deployment optimization of abnormal data adopts a triple verification mechanism, as follows: Parameter correlation verification: Parameter consistency is judged based on the Pearson correlation coefficient, as shown in the following formula:

[0036] The association is deemed valid; Sensor self-diagnosis: Monitoring sensor signal stability and coefficient of variation The sensor was determined to be normal. Indicates the sample standard deviation. This represents the sample mean; Historical data comparison: Calculate the deviation rate between the current data and the historical data for the same period. If:

[0037] Then it is determined to be valid, where, This is the average value of the parameter over a historical period. Example 2: The monitored object was an auxiliary transformer (model: TBQ35-800 / 25, rated capacity 800kVA), with the following operating conditions: traction load rate 40%-85% (average 62%), ambient temperature -25℃~40℃, cabinet vibration amplitude 0.1g~0.3g, and power grid fluctuation coefficient ≤±5%; The basic parameters of the adaptive fusion learning unit are as follows: Model architecture: Attention-LSTM (128 hidden layer dimensions, 4 attention heads) + XGBoost (200 decision trees, maximum tree depth 6). Weighting coefficients =0.6 (LSTM model weights), 1- =0.4 (XGBoost model weights); Training dataset: Factory rated parameters of the same model transformer: initial insulation resistance 1000MΩ, initial winding resistance 0.02Ω, temperature threshold 75℃; Individual historical data: Operating data of this locomotive over the past 2 years (cumulative 8000 hours), 3 minor fault records (2 instances of insulation dampness, 1 instance of winding overheating); Real-time data acquisition frequency: 5kHz (key parameters such as temperature and voltage), 1kHz (environmental and operating condition parameters); The health level thresholds are shown in the table below:

[0038] RUL algorithm parameters: safety factor =1.2, operating load factor =0.62 (converted from average load rate), Environmental Impact Factor =0.35 (Combined calculation of temperature, humidity and vibration), weighting coefficient =0.3、 =0.2; The iteration triggering conditions are as follows: Excellent or Good: Accumulated 300 hours of valid data or one instance of abnormal parameter fluctuation triggering an iteration; Medium or Poor: 50 hours of accumulated valid data or two consecutive non-correlated anomalies triggering an iteration; Danger: Real-time iteration (model parameters are updated every 10 minutes); Manual calibration weight: 30% (fault handling results and component test data entered by maintenance personnel); Parameter correlation verification: Pearson correlation coefficient ; Sensor self-diagnosis: coefficient of variation ; Historical data comparison: deviation rate ≤10%; Implementation process and results: During a certain period, the sensing system collected the following data: Real-time operating parameters: winding temperature 78℃, insulation resistance 280MΩ, winding resistance 0.023Ω (attenuation 15%), current 320A, voltage 25kV; Environmental and operating parameters: ambient temperature 32℃, cabinet vibration 0.25g, traction load rate 78%; Anomaly detection process: Parameter correlation verification: Calculate the Pearson correlation coefficient between winding temperature and current. =0.72 (≥0.6), the association is considered valid; Sensor self-diagnosis: 10 sets of temperature sensor data (76℃, 77℃, 78℃, 77℃, 79℃, 78℃, 79℃, 80℃, 78℃, 77℃), average value. =77.9℃, standard deviation =1.1℃, coefficient of variation =1.41% (≤5%), sensor is normal; Historical data comparison: The average insulation resistance for the same period in history was 305 MΩ, and the current value was 280 MΩ. The deviation rate was approximately 8.2% (≤10%), based on |280-305|÷305. The data is valid. Adaptive fusion learning unit diagnosis and lifetime prediction: Substituting into the formula for calculation, Input parameter vector at time step =[78,280,0.023,320,25,32,0.25,0.78];Hidden layer state After training =[0.12,0.08,...,0.05]、 =0.03, attention weight The final output is the key features. =0.35 × Temperature characteristic + 0.28 × Insulation resistance characteristic + 0.18 × Winding resistance characteristic + ... (Focusing on core parameters); Multi-model fusion diagnostics: LSTM model output =Insulation resistance decay + mild overheating (92% confidence level), XGBoost model output =Insulation aging + Overheating due to excessive load (confidence level 89%), fusion result =0.6×92%+0.4×89%=90.8%, final diagnosis: insulation resistance decayed to a medium level, accompanied by slight overheating, which is related to excessive load rate and accumulated ambient humidity; Remaining life prediction: Base value calculation: =100MΩ (allowable threshold for insulation resistance). =280MΩ, = (1000-280)÷(24×365×2)≈0.041MΩ / hour (average decay over the past 2 years), converted to monthly decay =0.041×24×30≈29.52MΩ / month =(100-280)÷(1.2×(-29.52))≈5.9 months (the negative sign indicates that the threshold has not been reached, so the absolute value is taken). Corrected lifespan: =5.9×(1-0.3×0.78-0.2×0.35)=5.9×(1-0.234-0.07)=5.9×0.696≈4.1 months; 3D output results: Quantitative value 4.1 months + health level "medium" + degradation rate 2.1% / month (monthly decay rate of insulation resistance); Iteration trigger: Model iteration is triggered after accumulating 50 hours of valid data due to the health level being "medium". Manual calibration: Maintenance personnel entered "Insulation oil was found to be damp during inspection; desiccant has been replaced; insulation resistance was retested at 310 MΩ" (manual verification data). ); Parameter update: =0.7× +0.3× After the update, the model's recognition weight for "insulation dampness" type faults increased by 15%, and the confidence level increased to over 95% when the same type of data was input again. The fault warning module triggered a level-two warning, which was pushed to the maintenance team via the vehicle-mounted APP. The warning message stated: "Auxiliary transformer health level 'Medium', remaining lifespan 4.1 months, degradation rate 2.1% / month; it is recommended to schedule a shutdown for maintenance within one month, focusing on replacing the insulating oil, optimizing the cooling air duct, and controlling the traction load rate to no more than 70%." Based on this recommendation, the maintenance team conducted a shutdown for maintenance before the deadline. After replacing the insulating oil, the insulation resistance recovered to 380MΩ, the health level rose to "Good", and the remaining lifespan prediction was updated to 8.3 months, verifying the accuracy of the system's diagnosis and prediction. Based on the above description, this invention uses Attention-LSTM+XGBoost multi-model fusion combined with attention mechanism to accurately extract key features, and improves fault diagnosis accuracy and reduces false alarm rate with triple abnormal data verification. It solves the problems of poor adaptability and high misjudgment rate of single algorithm. Furthermore, the RUL algorithm integrates multi-dimensional influencing factors, and the three-dimensional output mode clearly presents the equipment degradation status, providing solid data support for predictive operation and maintenance.

[0039] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. An online monitoring system for auxiliary transformers of heavy-haul railway locomotives, characterized in that: It includes a sensing system, a control system, and a monitoring center. The sensing system is communicatively connected to the control system, and the control system is communicatively connected to the monitoring center through a wireless communication network. The sensing system includes a temperature sensor, a thermal imaging sensor, a voltage sensor, a current sensor, a power quality monitoring sensor, a resistance measurement sensor, and an insulation resistance sensor. The sensing system integrates a data acquisition module and a data transmission module. The data acquisition module has multi-channel synchronous acquisition, high-speed sampling, and data caching functions. The control system includes a data analysis module and a fault early warning module. The data analysis module can perform data preprocessing, threshold judgment, trend analysis, fault diagnosis and data mining. The fault early warning module can issue early warning signals and record fault-related information through sound, light and text messages. The monitoring center includes a remote monitoring module and a data storage and management module. The remote monitoring module supports remote access, control, and debugging. The data storage and management module uses a MySQL or Oracle database and has data backup, recovery, statistics, and report generation functions.

2. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 1, characterized in that: The wireless communication network uses 5G wireless transmission technology, and the data transmission process is encrypted.

3. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 1, characterized in that: The sampling frequency of the data acquisition module is adapted to the monitoring requirements of the auxiliary transformer's operating parameters.

4. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 1, characterized in that: The data analysis module integrates an adaptive fusion learning unit based on Attention-LSTM+XGBoost.

5. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 4, characterized in that: The adaptive fusion learning unit extracts key parameter features through an attention mechanism, combines the gradient boosting tree algorithm to construct a multi-model collaborative diagnostic framework, and realizes quantitative assessment of the equipment's full life cycle based on the Remaining Life (RUL) prediction algorithm.

6. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 5, characterized in that: The training dataset of the adaptive fusion learning unit includes the factory rated parameters of transformers of the same model, individual historical operation-fault data, and real-time operation parameters. The remaining life prediction results are output in a three-dimensional mode, and the health level corresponds to a clear parameter threshold range. The life expectancy prediction results are output in a three-dimensional format as follows: quantified value + health level + decline rate. The quantified value is in monthly units, the health level includes excellent, good, moderate, poor, and critical, and the decline rate is in % / month.

7. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 6, characterized in that: The adaptive fusion learning unit adopts a hierarchical iterative optimization mechanism associated with health levels. Different health levels correspond to different iteration trigger conditions, and the iteration process incorporates manual verification data feedback calibration, with calibration accounting for 30% of the weight.

8. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 7, characterized in that: The data analysis module also integrates an intelligent verification unit for abnormal data, which filters invalid data through a triple mechanism of parameter correlation verification, sensor self-diagnosis, and historical data comparison.

9. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 8, characterized in that: The specific implementation process of the adaptive fusion learning unit is as follows: Based on LSTM neural networks, an attention mechanism is introduced to automatically assign parameter weights, focusing on parameters that play a key role in fault diagnosis and life prediction, including winding temperature and insulation resistance. The specific feature extraction formula is as follows: Hidden layer state update: Attention weight calculation: Weighted feature output: in, for Input parameter vector at any time, This represents the hidden state of the LSTM layer. For attention weight vectors, The final extracted key feature vector, , , , Parameters for model training; By integrating the XGBoost gradient boosting tree algorithm, a multi-model collaborative framework is constructed, and diagnostic results are output through a weighted voting mechanism, as shown in the following formula: Model output fusion: in, For the final fault diagnosis results, , These are the diagnostic outputs for the LSTM and XGBoost models, respectively. These are weighting coefficients; A quantitative model is constructed based on parameter decay patterns and historical fault data, with the following formula: Remaining lifespan base value: Corrected remaining lifetime: in, For the parameter, allowable threshold for Measured values ​​of time parameters This represents the average decay rate over the past three months. For safety reasons, This is the load factor for the operating condition, with a value between 0 and 1. This is the environmental impact coefficient, with a value between 0 and 1. , These are the weighting coefficients; The remaining life expectancy prediction results are output in three dimensions using a format of quantitative values ​​+ health level + decline rate.

10. The online monitoring system for auxiliary transformers of heavy-haul railway locomotives according to claim 9, characterized in that: Dynamically adjust the model iteration strategy based on the equipment health level: A health rating of Excellent or Good is achieved by accumulating 300 hours of valid data or triggering an iteration due to one instance of abnormal parameter fluctuation. Health level is medium or poor: 50 hours of valid data or two consecutive iterations triggered by unrelated anomalies; Health level is critical: Model parameters are updated iteratively in real time; During the iteration process, manually entered maintenance records and fault verification results are weighted at 30% for calibration, as shown in the following formula: in, For the updated model parameters, For iterative calculation of parameters, The calibration parameters were manually verified. The intelligent verification and deployment optimization of abnormal data adopts a triple verification mechanism, as follows: Parameter correlation verification: Parameter consistency is judged based on the Pearson correlation coefficient, as shown in the following formula: The association is deemed valid; Sensor self-diagnosis: Monitoring sensor signal stability and coefficient of variation The sensor was determined to be normal. Indicates the sample standard deviation. This represents the sample mean; Historical data comparison: Calculate the deviation rate between the current data and the historical data for the same period. If: Then it is determined to be valid, where, This is the average value of the parameter over a historical period.

Images